Claude 3.7 Sonnet (retired)

Claude 3.7 Sonnet introduced hybrid reasoning in February 2025 and was retired on 19 February 2026. Its original benchmarks, why it mattered, and what replaced it.

Why this page still exists: Claude 3.7 Sonnet is the hinge in Claude's version history, and a great deal of writing about AI reasoning from 2025 refers to it by name. People arrive here from old documentation, conference talks and blog posts, and the useful answer is what it was, what it started, and what to use instead.

Claude 3.7 Sonnet was released on 24 February 2025 and was the first Claude model with hybrid reasoning: a single model that could answer immediately or think visibly for as long as a developer allowed before answering. Every reasoning control in the current lineup - adaptive thinking, the effort ladder, thinking-token billing - descends from what shipped that day.

#Why Claude 3.7 Sonnet mattered

Before February 2025, the industry's answer to hard reasoning problems was a separate model: you routed easy questions to a fast model and hard ones to a slow "reasoning" model, and you had to decide which was which in advance. Claude 3.7 Sonnet collapsed that into one model with two modes. Anthropic's argument at the time was that reasoning is not a different product, it is a dial - and that the developer, not the router, should hold it.

That position won. Every current Claude model has thinking built in rather than bolted on, and on Opus 5, Sonnet 5 and Fable 5 it is on by default. What changed is who sets the depth: 3.7 Sonnet made you specify a thinking budget in tokens, whereas current models decide adaptively, guided by an effort parameter that runs low → medium → high → xhigh → max. The full evolution is documented on the adaptive and extended thinking page.

The other thing 3.7 Sonnet introduced was visible reasoning. Its thinking was returned as readable text, which made it possible to see where a model went wrong rather than only that it did. Current frontier models have retreated from that: thinking is returned summarized or omitted, and on Fable 5 the raw chain of thought is never returned at all. Whatever the safety merits, that is a genuine loss of debuggability compared with the February 2025 experience.

#Claude 3.7 Sonnet's benchmark record

These are February 2025 figures for a model that has been retired since 19 February 2026. They describe what the model did at launch and nothing about anything you can call today. Note also that SWE-bench Verified and GPQA are no longer reported by frontier vendors - Anthropic, OpenAI and Google have all moved to long-horizon agentic evaluations, so there is no honest way to line these numbers up against a 2026 model.

The prompt-injection figure is the one that aged best. It was early evidence that spending inference compute on deliberation makes a model harder to talk out of its instructions - an argument that now underwrites the entire agent security story, because an agent with tool access is a much larger target than a chatbot. If you are building anything that lets a model act on untrusted input, the practical patterns are on MCP and Claude agents.

The GPQA figure carries a caveat worth restating: 96.5% was reached with parallel test-time compute - sampling many independent attempts and selecting among them - not by one model turning in one answer. It measured what the model could reach with a compute budget, not what a single call would return. That distinction is still routinely blurred in benchmark reporting today.

#Two things that started with 3.7 Sonnet

#Where Claude 3.7 Sonnet fell short

It is worth being honest about the model rather than treating it as a milestone only. Two complaints recurred through 2025.

First, over-thinking. With a generous budget, 3.7 Sonnet would reason at length about questions that did not need it, burning output tokens and latency for no gain. Because the budget was a fixed number you set in advance, the model could not tell an easy question from a hard one - which is precisely the problem adaptive thinking was built to solve.

Second, over-eagerness in agentic settings. On long coding tasks it had a documented tendency to take shortcuts, edit more than asked, or declare success early. Anthropic explicitly claimed a large reduction in that behaviour when Claude 4 shipped three months later, which tells you how real the complaint was. If you are still writing prompts in a 2025 style, the prompt engineering page covers what current models actually need - notably that verification instructions carried over from older models now cause over-verification rather than fixing anything.

#Migrating away from Claude 3.7 Sonnet

Sonnet 5 costs $2 / $10 per million tokens against 3.7 Sonnet's launch price of $3 / $15, carries a 1M-token context instead of 200k, and emits roughly 30% more tokens for the same text under its newer tokenizer - so re-measure rather than assuming a straight saving. Plan and API rates are on the pricing page, and every currently callable model is listed on the model index. The next generation's archive is on the Claude Sonnet 4 page.

#Frequently asked questions